LangGraph

Author: Ludovic HENRY ludovic.henry@qti.qualcomm.com
Date: 2026-06-17
Scope: RISC-V (riscv64/linux) support status for LangGraph
Audience: Technical leadership, resource allocation strategy
Verification policy: Every claim is cross-referenced to a primary upstream source. Items that could not be verified against a second source are marked [NEEDS VERIFICATION].

1. Project Overview

LangGraph is a Python (and TypeScript) framework for building stateful, multi-actor agent workflows using a graph-based execution model. It sits on top of LangChain Core and provides cycle-aware directed graphs where nodes are Python callables and edges encode control flow. It does not contain LLM inference, compute kernels, or numeric routines - all heavy computation is delegated to downstream libraries (LLM APIs, vector stores, serialization libraries) and ultimately to cloud or hardware inference endpoints.

Governance: LangGraph is a fully corporate-controlled project. There is no foundation affiliation (not Linux Foundation, Apache, CNCF, or any equivalent). LangChain Inc. employs all core maintainers, including Harrison Chase (co-founder/CEO), William FH (@hinthornw), Vadym Barda (@vbarda), Eugene Yurtsev (@eyurtsev), Sydney Runkle (@sydney-runkle), Nuno Campos (@nfcampos), and David Duong (@dqbd). There is no MAINTAINERS, OWNERS, or CODEOWNERS file; maintainership is entirely internal to LangChain Inc.

License: MIT.

Repository created: 2023-08-09.

Community stance on new ports: Not applicable. Because LangGraph is a pure-Python package with no compiled components of its own, there is no port concept. The library installs on any Python-supported architecture via pip install langgraph without modification. No platform-acceptance policy exists and none is needed.

RISE membership: LangChain Inc. is not a RISE member. RISE Premier members are Alibaba, Google, MediaTek, NVIDIA, Qualcomm, Red Hat, SiFive, and Tenstorrent. General members are Akeana, Andes Technology, ESWIN, BISCOE, Canonical, Douyin Vision, ISCAS, Microchip, NextSilicon, SpacemiT, and ZTE. None are affiliated with LangChain.

2. Port History and Upstreaming Timeline

Date Event Source
2023-08-09 Repository created langchain-ai/langgraph
N/A First RISC-V-related commit None found
N/A First RISC-V CI job None found
N/A First riscv64 binary or wheel None found

No RISC-V port work has ever been done in this repository. GitHub searches for riscv repo:langchain-ai/langgraph and riscv64 repo:langchain-ai/langgraph return only false positives: uv.lock lock files that contain riscv64 wheel filenames for transitive dependencies (charset-normalizer, pydantic-core, ruff, ty) as PyPI distribution metadata. These are not LangGraph code or LangGraph CI. The one PR with “riscv” in its title (PR #7472) is a Dependabot uv version bump where “riscv” appears only inside the embedded uv release notes listing uv-riscv64gc-unknown-linux-gnu.tar.gz as a download artifact.

No community contributor has filed a RISC-V tracking issue, sent a RISC-V patch, or raised a RISC-V question in any issue or PR.

3. Upstream Support Tier

LangGraph has no documented platform tier policy. There is no PLATFORMS.md, SUPPORT.md, or tiering matrix of any kind.

In practice, the project implicitly supports any platform where Python runs, because its release artifacts are py3-none-any wheels - universal pure-Python wheels with no architecture discriminator.

The exception is langgraph-cli, the command-line tool for building and deploying LangGraph Cloud containers. Its deploy.py and docker.py source files hard-code linux/amd64 as the sole Docker target platform. This means the container deployment workflow for LangGraph Cloud is amd64-only. The library itself is not affected.

Component amd64 arm64 riscv64
langgraph (Python library) Yes - pip universal wheel Yes - pip universal wheel Yes - pip universal wheel
langgraph-checkpoint Yes - pip universal wheel Yes - pip universal wheel Yes - pip universal wheel
langgraph-sdk Yes - pip universal wheel Yes - pip universal wheel Yes - pip universal wheel
langgraph-prebuilt Yes - pip universal wheel Yes - pip universal wheel Yes - pip universal wheel
langgraph-cli container deploy Yes - explicitly targeted No - hard-coded linux/amd64 excludes arm64 No - hard-coded linux/amd64 excludes riscv64
Official CI (GitHub Actions) Yes - ubuntu-latest (x86_64) No No
Release-blocking CI Yes (x86_64) No No

4. Technical Architecture and RISC-V-Specific Subsystems

LangGraph has no architecture-specific subsystems. The project contains:

  • No C or C++ source files
  • No assembly files
  • No JIT compiler
  • No SIMD dispatch layer
  • No cryptographic primitives
  • No garbage collector barriers
  • No numeric kernels
  • No arch/riscv/ or equivalent directory

A full code search for vfloat32m1_t, rvv, riscv, __riscv, _mm256, __ARM_NEON, and related architecture intrinsic patterns returns zero genuine hits in LangGraph source code.

Component amd64 arm64 riscv64
Graph execution engine Pure Python Pure Python Pure Python
State management / checkpointing Pure Python Pure Python Pure Python
Serialization (JSON/msgpack) Delegated to orjson/ormsgpack Delegated to orjson/ormsgpack Delegated to orjson/ormsgpack
Data validation Delegated to pydantic-core Delegated to pydantic-core Delegated to pydantic-core
TypeScript/JS SDK Pure TypeScript Pure TypeScript Pure TypeScript

All performance-sensitive operations in LangGraph are delegated to dependency libraries (see Section 9). LangGraph itself has no RISC-V gap in its own code because it has no architecture-specific code at all.

5. Build System, Cross-Compilation, and Toolchain

LangGraph has no native build system. There is no CMake, Meson, autoconf, Makefile, Cargo workspace, or any native compilation step in langchain-ai/langgraph.

Build and release use standard Python tooling:

  • uv for dependency management and lock files
  • pyproject.toml per sub-package
  • maturin is used only by dependency libraries (pydantic-core, orjson, ormsgpack), not by LangGraph itself
  • GitHub Actions release.yml publishes to PyPI via standard uv publish

For riscv64 specifically: pip install langgraph on a riscv64 Linux system downloads the py3-none-any universal wheel and installs it without compilation. No Rust toolchain, no C compiler, and no QEMU are required to install or run LangGraph itself.

Cross-compilation: not applicable to LangGraph. Dependencies with compiled extensions (pydantic-core, orjson, ormsgpack) require a Rust toolchain when building from source on riscv64, since no prebuilt riscv64 wheels exist for those packages (see Section 9).

6. Feature Coverage and Gap Analysis vs arm64 and amd64

Feature amd64 arm64 riscv64 Gap
Graph execution Full Full Full None
State persistence (checkpoint-sqlite) Full Full Full (sqlite-vec C ext builds from source [NEEDS VERIFICATION]) None known
State persistence (checkpoint-postgres) Full Full Full (psycopg from source; psycopg-binary not available) Minor: no binary wheel for psycopg
JSON serialization (orjson) Full (prebuilt wheel) Full (prebuilt wheel) Source build only, no prebuilt wheel Build friction only
Msgpack serialization (ormsgpack) Full (prebuilt wheel) Full (prebuilt wheel) Source build only, no prebuilt wheel Build friction only
Pydantic validation (pydantic-core) Full (prebuilt wheel) Full (prebuilt wheel) Source build only, no prebuilt wheel Build friction only
LangGraph Cloud container deploy (CLI) Full Not supported Not supported Hard-coded linux/amd64 in CLI
Async event loop (uvloop, optional) Prebuilt wheel available Prebuilt wheel available Source build only; PR #733 closed without merge Not a runtime dep; asyncio fallback works

Functional gaps: none in the core library. The langgraph-cli container deployment tooling is blocked on amd64 only.

Performance gaps: Data not available - no published benchmarks comparing LangGraph performance on riscv64 vs arm64 or amd64 exist anywhere. Given that LangGraph performs no compute itself (all serialization is in delegated deps), any performance delta would be in orjson and pydantic-core serialization paths, which incur source-build overhead but no functional penalty.

Security hardening gaps: Data not available - no security hardening analysis for riscv64 (stack canaries, CFI, pointer authentication equivalents) was found in any upstream source.

NaN / floating-point issues: Data not available - not applicable. LangGraph performs no floating-point operations.

7. CI/CD Infrastructure

All 17 GitHub Actions workflow files in .github/workflows/ were read directly from the repository. None contain any reference to riscv, riscv64, or linux/riscv64. The full list of workflow files checked:

_integration_test.yml, _lint.yml, _sdk_integration_test.yml, _test.yml, _test_langgraph.yml, _test_release.yml, baseline.yml, bench.yml, ci.yml, deploy-redirects.yml, pr_lint.yml, release.yml, reopen_on_assignment.yml, require_issue_link.yml, tag-external-issues.yml, tag-external-prs.yml, uv_lock_ugprade.yml

No QEMU setup action, no runs-on: [self-hosted, riscv64], no matrix entry with arch: riscv64, and no RISE runner reference appear in any file.

CI dimension amd64 arm64 riscv64
Unit tests Yes - ubuntu-latest No No
Integration tests Yes - ubuntu-latest No No
Lint Yes - ubuntu-latest No No
Release pipeline Yes - ubuntu-latest No No
Hardware used GitHub-hosted ubuntu-latest (x86_64) N/A N/A
RISE runners No No No

No .gitlab-ci.yml, Jenkinsfile, or .cirrus.yml are present in the repository.

8. Distribution and Release Status

PyPI: Latest version is 1.2.11. Every published file is either langgraph-X.X.X-py3-none-any.whl or langgraph-X.X.X.tar.gz. The py3-none-any tag is the universal pure-Python wheel tag; no platform-specific builds exist. This has been consistent across all releases from 0.0.8 through 1.2.11. Source: PyPI JSON API for langgraph.

GitHub Releases: Three most recent releases (1.2.11, checkpointpostgres==3.1.2, checkpoint==4.2.0) have zero attached binary assets. LangGraph publishes exclusively to PyPI.

Ubuntu 24.04 Noble: LangGraph is not packaged. Source: Ubuntu package search returned no results.

Debian: LangGraph is not packaged. https://tracker.debian.org/pkg/langgraph returns HTTP 404.

Arch Linux RISC-V (archriscv.felixc.at): LangGraph is absent. Search returned no results.

RISE wheel builder: LangGraph is absent from the RISE Python wheel builder package list. The index redirects to standard PyPI.

What a user must do to get a working installation on riscv64:

  1. Ensure Python 3.9+ is available on the riscv64 system.
  2. Run pip install langgraph. The universal wheel installs without compilation.
  3. For dependencies with native extensions (pydantic-core, orjson, ormsgpack, xxhash), pip will attempt to build from source if no prebuilt riscv64 wheel is available. This requires a Rust toolchain (for pydantic-core, orjson, ormsgpack, uuid-utils) and a C compiler (for xxhash, sqlite-vec if used).
  4. Build time for pydantic-core from source on a low-end RISC-V board is approximately 15 minutes [NEEDS VERIFICATION - this figure comes from a single source in the pydantic-core issue tracker].
  5. psycopg (not psycopg[binary]) must be used for PostgreSQL checkpointing; the binary wheel is unavailable.

9. Dependencies

Summary Table

Dependency Role riscv64 build riscv64 test riscv64 prebuilt wheel Status
langchain-core Graph node protocol, message types Pure Python N/A py3-none-any No gap
langgraph-checkpoint State persistence base Pure Python N/A py3-none-any No gap
langgraph-sdk REST API client Pure Python N/A py3-none-any No gap
xxhash Node state hashing Yes - C ext builds No dedicated CI Yes - official manylinux riscv64 wheels since v4.0.0 No gap
pydantic-core Pydantic v2 validation (Rust/PyO3) Yes - builds from source No CI for riscv64 No prebuilt wheel Build friction; ~15 min source build [NEEDS VERIFICATION]
uuid-utils UUID generation (Rust/PyO3) Yes - builds from source No CI No prebuilt wheel Build friction
orjson Fast JSON serialization (Rust/PyO3) Yes - builds from source No CI for riscv64 No prebuilt wheel Build friction
ormsgpack MessagePack serialization (Rust/PyO3) Yes - builds from source No CI for riscv64 No prebuilt wheel Build friction
sqlite-vec Vector similarity (checkpoint-sqlite, optional) Likely - C ext No CI for riscv64 No prebuilt wheel Unverified on riscv64 [NEEDS VERIFICATION]
psycopg PostgreSQL driver (checkpoint-postgres, optional) Yes CI added (PR #1197 merged) No binary wheel; psycopg (not psycopg[binary]) works Minor gap: binary extra unavailable; tests failing tracked in issue #883
uvloop High-performance async loop (optional dev/test dep only) Yes - builds from source No CI No prebuilt wheel; PR #733 closed without merge Not a runtime dep; asyncio fallback works

Notable dependency details

pydantic-core: Issue #1906 was closed 2026-03-11 with a reference to pydantic PR #12723 (merged 2026-02-25, titled “Add riscv64 build target for manylinux”). Despite this PR, riscv64 is not yet present in the pydantic-core CI matrix (which covers x86_64, aarch64, armv7, i686, ppc64le, s390x). No riscv64 wheel is published to PyPI. Source builds work but require the Rust toolchain and significant build time.

orjson: Issues are disabled on the orjson repository. The CI matrix covers x86_64, aarch64, armv7, i686, ppc64le, and s390x. riscv64 is absent. No tracking issue exists. Source build via Rust/maturin functions but no prebuilt wheel is published.

xxhash: Full riscv64 support as of v4.0.0. Official manylinux and musllinux riscv64 wheels are published covering Python 3.9-3.15. This dependency is fully unblocked.

psycopg: PR #1197 merged, adding riscv64 CI. Issue #883 tracks failing tests on riscv64. The psycopg[binary] extra (pre-compiled C driver) has no riscv64 wheel. The psycopg package (pure Python with C driver compiled at install time) works on riscv64 and is the correct choice for production use.

11. Known Bugs and Active Issues

| ID | Title | Status | Severity | Notes | |—-|——-|——–|———-|——-| | - | - | - | - | - |

No riscv64-related bugs, correctness issues, or performance reports exist in langchain-ai/langgraph. GitHub issue searches for all RISC-V query variants returned zero results. LangGraph has no architecture-specific code paths that could exhibit riscv64-specific behavior.

12. Objections and Upstream Blockers

Stated objections: None found. No maintainer or contributor has made a statement about RISC-V in any issue, PR, or discussion in the repository.

Technical blockers for the library: None. LangGraph itself is unblocked on riscv64.

Technical blockers for the dependency stack:

  • pydantic-core: no prebuilt riscv64 wheel; source build required; issue closed but wheel publishing not yet implemented
  • orjson: no prebuilt riscv64 wheel; no tracking issue; issues disabled on the repo
  • ormsgpack: no prebuilt riscv64 wheel; no tracking issue
  • psycopg: tests failing on riscv64 (issue #883); binary extra unavailable

Organizational blockers: None identified. LangChain Inc. is a venture-backed startup with no formal platform support policy. Upstream maintainers of dependencies (pydantic, orjson, psycopg) operate independently.

Acceptance probability for RISC-V patches: High for pure build/CI additions to dependency libraries (pydantic-core, orjson), given that these libraries already support ppc64le and s390x as precedent platforms. No political opposition to riscv64 has been stated. The main blocker is engineering capacity in those upstream projects, not policy.

13. Investment Analysis

RISE has done no reported work on LangGraph or its direct dependency chain (as of the RISE blog scan through 2026-07). RISE AI/ML Working Group activity is focused on compiler and inference runtime work (IREE, XNNPACK, MLIR), not Python orchestration frameworks.

13.1 Functional Enablement

LangGraph itself requires zero functional enablement work for riscv64. The library is pure Python and installs via pip today. The functional gaps are entirely in third-party dependency libraries (pydantic-core, orjson, ormsgpack, psycopg).

13.2 Performance Optimization

Data not available - no benchmark data exists for LangGraph on any architecture. LangGraph performs no compute itself; its per-invocation overhead is Python graph traversal and serialization calls to orjson/ormsgpack/pydantic-core. Performance optimization of LangGraph on riscv64 is not meaningful without first establishing baseline measurements.

13.3 CI/CD Infrastructure

LangGraph has no riscv64 CI. Adding riscv64 CI to LangGraph itself would be low-cost but low-value: since LangGraph is pure Python, a CI job would trivially pass and provide no signal about architecture-specific correctness. The meaningful CI investment is in the dependency libraries (pydantic-core, orjson, ormsgpack), where compiled extensions exist and architecture-specific failures are plausible.

13.4 Ecosystem Enablement

The primary ecosystem enablement gap is the absence of prebuilt riscv64 wheels for pydantic-core, orjson, and ormsgpack. This forces Rust toolchain presence and long source builds on any riscv64 deployment system, which is a significant friction point for production deployments.

13.5 Summary Table

Area Work Item Effort (person-weeks) Owner Priority
Functional Add riscv64 to pydantic-core CI and wheel build matrix 2 pydantic-core upstream / RISE High
Functional Add riscv64 to orjson CI and wheel build matrix 2 orjson upstream / RISE High
Functional Add riscv64 to ormsgpack CI and wheel build matrix 1 ormsgpack upstream / RISE Medium
Functional Resolve psycopg riscv64 test failures (issue #883) 2 psycopg upstream Medium
Functional Add riscv64 to uvloop wheel build matrix (issue #732) 1 uvloop upstream Low (not a runtime dep)
CI/CD Add riscv64 to langgraph-cli Docker targets (deploy.py, docker.py) 1 LangChain Inc. Low (cloud product, not core library)
CI/CD Add riscv64 smoke-test job to langgraph GitHub Actions 0.5 LangChain Inc. Low (pure Python; low signal value)
Performance Establish riscv64 baseline benchmarks for LangGraph serialization paths 1 RISE / Qualcomm Low (no compute in library itself)

Total estimated effort: approximately 10.5 person-weeks, almost entirely in upstream dependency libraries, not in LangGraph itself.

Priority assessment: For a chip company evaluating RISC-V investment in the AI agent framework space, LangGraph itself presents no investment barrier - it works on riscv64 today. The real investment target is the compiled extension layer (pydantic-core, orjson), which blocks frictionless deployment of the broader Python AI stack on riscv64. Those efforts have value independent of LangGraph and should be tracked at the pydantic and orjson project level, not the LangGraph level.

14. Updates

No updates yet - initial report dated 2026-06-17.

15. References